| Shoeprint image is one of the most important evidences in criminal investigation,It plays a major role in exposing and corroborating crime.So far,most algorithms for shoeprint classification are based on closed-set.It assumes that all test images contain categories that have appeared in the training set.However,in the real scene,there will be many unknown classes of shoeprints,if we use the assumption of closed set directly.Its closed set will force the classifier to choose from known classes.Classification performance is not good.In recent years,deep networks have made significant progress in solving various visual recognition problems.It has a huge impact on academic and commercial applications.Deep network can extract deep features from the input shoeprint image.Based on this,this paper uses the deep learning technology to propose the research on open set shoeprint classification algorithm.The main work is as follows:1)An open set classifieation algorithm based on fusion multi-convergence layer is presented.By analyzing the characteristics of shoeprint images,based on the deep network,this paper gives a method for feature extraction on shoeprint images.Because there are many categories to be classified,method of merging convolution response maps of different layers.Combine the global information of the niddle layer and the local information of the latter layer to improve the distinguishability between categories.And combine threshold method,accomplishing the open set classification.The AUC on the shoeprint dataset reaches 0.84.2)An open set classification algorithm based on multi-scale featxure weighted fusion is presented.This article focuses on the characteristics of the shoeprint and the heel,An open set classification algorithm based on multi-scale feature weighted fusion is presented.This article cuts a shoe print image into the soles of the feet,the center of the soles,and the heels.Deep network extraction features using the sane network structure for each part.After normalizing each feature,a weighted fusion method is used to obtain a feature representation of a shoeprint image.Combine the threshold method to accomplish open set classification.Experiments show that the algorithm can effectively increase the distance between classes and reduce the distance within the class.The AUC on the shoeprint image dataset reaches 0.89.3)An open set classification algorithm based on statistical extreme theory is presented.This paper addresses the problem in the traditional open set classification,that is,the a priori assumption that the threshold method is directly used is that the category of the test sample must appear in the training set.The method of using thresholds does reject some out-of-set images,but the threshold method is not optimal.Open set classification algorithm based on statistical extreme theory can not only improve the performance of open-collection classification,but also reduce the risk of open space.Experiments show that the algorithm is significantly improved in accuracy and AUC compared with the threshold method.The AUC on the shoeprint dataset reaches 0.91. |
